How to choose your first AI use case

The first AI project in a company carries more weight than it should. If it works, the second and third get funded easily. If it drifts, ’we tried AI’ becomes a reason not to try again for two years. So the choice matters, and the right choice is usually less exciting than the one people want.

Four questions we score every idea on

When we map opportunities with a client we typically collect thirty to forty ideas. Each one is scored on the same four questions.

• Value. If this worked perfectly, which number moves, and by how much? Hours, error rates, days of delay, revenue. If nobody can name the number, the idea is not ready.
• Feasibility. Is the task one that current models do reliably, such as reading, classifying, summarising, drafting and matching? Or does it need judgment that even your experts disagree on?
• Data. Do examples of the work exist, in volume, in a form you can get at? Two thousand past cases beat a perfect specification.
• Risk. What does a mistake cost, and how quickly would you notice? A wrong draft that a person reviews is cheap. A wrong payment is not.

What good first projects have in common

Looking back over the first projects that went well for our clients, a pattern is clear. They were high-volume and repetitive, hundreds of cases a week and not a dozen. They had an owner, one named manager who wanted the problem gone. A person stayed in the loop, so early errors were caught and cost little. And they touched one or two systems, not seven.

Document intake, triage of incoming requests, drafting replies from known information and searching internal knowledge come up again and again. They are not glamorous. They work.

What to leave for later

Anything customer-facing with no human review. Anything that depends on data you plan to clean up first. Anything described as a platform. Anything whose main sponsor is the vendor. These can all be good second or third projects, once you have a team that has shipped one.

A quick test

Before committing, try to finish this sentence: ’Today, task X takes our team Y hours a week with a Z percent error rate, and in six months it will take...’ If you can complete it with numbers and a name beside it, you have found your first use case. If you cannot, the first project is finding those numbers.